Datasets:

xiaoguangwang
/

syzfix-dataset

language:
  - en
tags:
  - linux-kernel
  - syzbot
  - bug-fixing
  - program-repair
  - security
pretty_name: SyzFix
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
dataset_info:
  features:
    - name: bug_id
      dtype: string
    - name: title
      dtype: string
    - name: crash_report
      dtype: string
    - name: c_reproducer
      dtype: string
    - name: syz_reproducer
      dtype: string
    - name: fix_commit_hash
      dtype: string
    - name: fix_commit_message
      dtype: string
    - name: final_patch_diff
      dtype: string
    - name: patch_evolution
      dtype: string
    - name: subsystem
      dtype: string
    - name: first_crash_date
      dtype: string
    - name: fix_date
      dtype: string
    - name: num_patch_versions
      dtype: int64
    - name: has_discussion
      dtype: bool
  splits:
    - name: train
      num_bytes: 5452023509
      num_examples: 7208
  download_size: 2211556912
  dataset_size: 5452023509

SyzFix — Fixed Linux Kernel Bugs, Full Lifecycle

7,000+ fixed Linux kernel bugs from syzbot, each captured end-to-end: the raw crash report, syzkaller/C reproducers, every patch revision posted to lore.kernel.org with its review discussion, and the final commit merged into torvalds/linux.

Code & docs: https://github.com/sysec-uic/syzfix · Website: https://sysec-uic.github.io/syzfix/

What's in the repo

Path	Contents
`data/train-*.parquet`	The structured dataset — one record per bug (schema below)
`processed/processed.jsonl.gz`	Full per-bug JSON (~2 GB gz, ~11 GB unpacked): complete mailing-list threads, all crash variants, raw syzbot fields

Record schema (main dataset)

Field	Description
`bug_id`, `title`, `status`	syzbot bug identity
`crash_report`	Full kernel oops / KASAN / BUG output
`c_reproducer`, `syz_reproducer`	Reproducers when available
`fix_commit_*`	Hash, title, author, date of the merged fix
`final_patch_diff`	The merged patch
`patch_evolution`	JSON: v1 → v2 → … diffs with inline discussions
`num_patch_versions`, `has_discussion`	Convenience flags

Usage

from datasets import load_dataset

ds = load_dataset("xiaoguangwang/syzfix-dataset")   # the structured dataset
print(ds["train"][0]["title"])

Full per-bug JSON (for custom processing / re-deriving training tasks):

pip install "syzfix @ git+https://github.com/sysec-uic/syzfix"
python -m dataset.restore_processed --repo xiaoguangwang/syzfix-dataset

Updates

The dataset tracks syzbot continuously; it is refreshed by an incremental crawler (python -m dataset.update in the GitHub repo).

Sources & license

Aggregates publicly available content from syzkaller.appspot.com, lore.kernel.org, git.kernel.org and patchwork.kernel.org. Kernel patches and mailing-list text remain under their original licenses/copyrights; the collection tooling is MIT-licensed. Intended for research use.

Citation

Please cite the GitHub repository (https://github.com/sysec-uic/syzfix) until the accompanying paper is available.